Trust state

A trust state is the label on a record that says how far it has been checked, separating raw evidence you just saved from a processed copy and from a reviewed record you are willing to act on.

A trust state is the label on a record that says how far it has been checked, separating raw evidence you just saved from a processed copy and from a reviewed record you are willing to act on. Each level carries a different promise about reliability. Say you record a phone call about a password reset: the audio file is raw, an AI transcript of it is processed, and the short list of confirmed steps you read and approved is reviewed. Tell the assistant to tag every record with its current state and to keep raw, processed, and reviewed clearly apart, so a draft built from unchecked material never gets treated as settled fact.

Builder example

When an assistant retrieves your records to answer a question, it has no way to weight them unless each one declares how far it was checked. A support bot that quotes an unverified transcript as if it were an approved policy can repeat a guess to a customer. Tell the system to store a trust state on every record and to prefer reviewed records when it answers, falling back to raw material only when it flags that the claim is unconfirmed.

Common confusion: A trust state is not the same as how recent a record is. A freshly captured note can be brand new and still unverified, while an older record can be reviewed and dependable. What the trust state tracks is the level of checking a record has passed, not its age.